Context before conclusions.
Writing is shaped by genre, editing, language background, and the purpose of a document. A short application statement and a technical résumé should not be interpreted in the same way.
Our reporting framework keeps three things distinct: observable features of the text, a reviewer’s editorial judgment, and evidence about the writing process. None should be silently substituted for another.
Make the method visible.
| Assessment type | What it means | Current website |
|---|---|---|
| Human review | A reviewer evaluates expression and explains their observations using excerpts.Illustrative sample report||
| Descriptive statistics | Local calculations of words, sentences, and average sentence length.Not offered on this showcase website||
| Automated AI detection | A trained detector’s output, with its own model, validation, and limitations.Not connected in this release||
| Writing-process evidence | Version history and documented revisions may provide context about how a text developed.Not collected on this website
What the AI / Human split means.
In a human-review report, AI-like expression is an estimate entered by a reviewer. It concerns the reviewer’s impression of expression; it does not measure the proportion of words generated by AI.
Human expression is the complementary display value: 100 minus the AI-like expression estimate. It is not an independently measured probability.
Example: a reviewer-entered value of 5 produces an AI-like expression display of 5% and a Human expression display of 95%. A value of 0 produces 0% and 100%. An unfilled assessment should remain “Not assessed,” never default to zero.
A 0% value does not prove that AI was never used. A 100% value does not prove that AI was used. These estimates are not plagiarism scores, certificates of original authorship, or promises about another platform’s results.
Automated detectors also require careful interpretation. Turnitin, for example, describes its percentage as the share of qualifying text its model flags, and suppresses exact values in the 1–19% range because of false-positive concerns. This reference is provided for context; DraftVerity is not affiliated with or endorsed by Turnitin. Read Turnitin’s explanation.
Keep findings tied to the passage.
- Identify the text. A useful issued report should name the document and the version being assessed.
- Quote faithfully. An excerpt should reproduce the source text, with enough context to make the observation understandable.
- Separate observation from inference. “Three sentences begin with the same phrase” is observable. “AI wrote this paragraph” requires different evidence.
- Make suggestions specific. Describe what a revision could clarify rather than prescribing a desired detector score.
- Retain uncertainty. Missing context should be acknowledged. Personal details in a passage are not verification that the described events occurred.
A showcase of the framework.
This website presents the reporting approach and an illustrative sample. It does not accept text or document uploads, run public analyses, or generate reports from a visitor’s content.
Sign-in and invitation registration forms demonstrate member access. In this preview, they do not contact an account service, verify invitations, or create accounts. An invitation code is required to submit the registration preview.
Every completed sign-in submission displays “User not found” as a clearly labeled preview response. No account lookup is performed, and submitted values are not sent or stored.
Explore the sample reportA transparent starting point.
This website introduces the DraftVerity reporting framework and an illustrative report. Live report issuance, identity verification, a public verification registry, and a trained AI-detection service are not available in this release.
The sample document, its commentary, its identifiers, and its percentages are illustrative. No university acceptance, institutional certification, detection accuracy, or independent accreditation is claimed.